How to Build Effective Product Loops in Claude Code | Tyler Folkman | Chief AI Officer, JobNimbus

Everybody says loops are the new prompts. Almost nobody explains what that means for product work. Tyler Folkman, Chief AI Officer and Head of Product at JobNimbus, builds a loop live on screen, closes it, and shows the exact hooks, gates, and skills his product, design, and engineering teams run every day. Full Writeup: https://www.news.aakashg.com/claude-code-loops Transcript: https://tinyurl.com/ytr6cw87 Timestamps 0:00 - Intro 1:02 - Intro 2:22 - The system behind daily loop engineering 7:53 - What turns a skill into a loop 10:04 - Ads 12:13 - The four parts of a working loop 16:40 - Generating prototype variants live 22:31 - AI built onboarding and docs written for AI 28:36 - Building a non slop loop from scratch 31:47 - Ads 35:56 - Running the decide skill and closing the loop live 42:47 - The loops and hooks every PM should build 49:03 - Quality loops and whether PMs should push code 58:32 - Design loops and the rise of the product builder 1:07:44 - Outro 🏆 Thanks to our sponsors: 1. Customer.io: Send smarter messages using your product data - http://customer.io/productgrowth 2. Ariso: Ship AI agents and features faster with fewer regressions - https://ariso.ai/aakash 3. Viktor: AI employee connected to 3k+ tools with every action approved by your team - http://viktor.com/aakashgupta3 4. Bolt: Build a complete design system from your code - http://bolt.new/aakash 5. Product Faculty: Get $150 off their #1 AI PM Certification with code AAKASH150 - https://www.productfaculty.com/?code=AAKASH150 Key Takeaways 1. A skill becomes a loop when you feed the learning back - A static skill sits there and stays the same forever. The loop closes when you hand the AI the log of the whole session and ask it what to improve. Skip that step and you are just running a skill. 2. The gate is the most important part of the loop - Fetch inputs, do work, pass a gate, write the artifact. The gate is where you validate the work was correct, and making it deterministic matters more than anything else in the chain. 3. Agents make decisions, prompts wait for you - The difference is not the model. An agent gets a task plus a way to verify itself and executes as far as it can alone. Prompting keeps you in the loop, which caps how far you can scale yourself. 4. Write the first version of a skill by hand - Human authored skills tend to beat AI authored ones because you know more about what you actually want. Once you are on the AI loop it moves fast enough that injecting your own thinking gets hard, like promising you will still pedal on an ebike. 5. In product, the gate is a customer, and that breaks the loop - Code loops run fast because the gate is a test. You cannot lock customers in a room and iterate on them. The workaround is mining the research calls and transcripts you already have to build a cheap first filter. 6. Prototypes are free now, so generate variants instead of one answer - The internal standard is at least three variants per idea. One minimal, one full featured, one creative. Then narrow a hundred ideas to five before anyone talks to a real customer. 7. Synthetic customers filter, real customers decide - Customer research transcripts loaded into a warehouse let AI inspect a prototype as your customer. It is not a high bar, but it catches low hanging failures fast, which is the whole point of a gate. 8. Write docs for AI, keep human docs to three pages - Part of onboarding is not written for people to read. It is context so the AI can answer questions. Anything a human is expected to read should be one to three pages, visual, and cut down by hand. 9. Hooks add determinism that a prompt cannot - Telling Claude never to delete everything or never to share credentials only works if it reads that instruction today. A hook fires on the bash command itself. Session close hooks can also force the improvement step you would otherwise forget. 10. Ship AI output you have not thought about and it costs you - Passing up unreviewed AI work just moves the effort to someone busier than you. Answering a question in a meeting with what Claude said is the failure mode. Use AI to push your thinking, not to replace the part where you do it. 👨‍💻 Where to find Tyler Folkman LinkedIn: https://www.linkedin.com/in/tylerfolkman X: https://x.com/tyler_folkman Website: https://www.tylerfolkman.ai/ Newsletter: https://tylerfolkman.substack.com/ 👨‍💻 Where to find Aakash X: https://x.com/aakashgupta LinkedIn: https://www.linkedin.com/in/aagupta/ Newsletter: https://www.news.aakashg.com/ #AIProductManagement #ProductManagement #ClaudeCode 🧠 About Product Growth The world's largest podcast focused solely on product + growth, with over 200K+ listeners. 🔔 Subscribe and turn on notifications.

Tyler FolkmanguestAakash Guptahost

Sep 4, 20261h 8mWatch on YouTube ↗

WHAT IT’S REALLY ABOUT

Building self-improving AI loops for product, design, and engineering work

A “loop” is a skill that fetches inputs, does work, passes a validation gate, produces an artifact, and then learns from logs to improve itself over time.
Tyler demonstrates a terminal-based orchestration workflow (via Herder and Claude skills) to run multiple agents, generate prototype variants live, and manage iterative improvement with Git and structured feedback.
The hardest part of product loops is defining strong gates—especially when the gate is real customer validation—so JobNimbus combines synthetic feedback from mined research with frequent real customer conversations.
To avoid “AI slop,” Tyler recommends human-authored skills, concise human-facing docs, AI-written docs for AI consumption, and hooks that enforce safety/quality deterministically rather than relying on prompt compliance.
As AI increases shipping velocity, engineering must prioritize “quality loops” (tests, standards checks, CI/CD) so customers don’t experience a higher absolute number of defects, while roles converge toward cross-functional “product builders.”

IDEAS WORTH REMEMBERING

5 ideas

A loop becomes a loop only when it learns.

Running a repeatable skill is useful, but it’s not a compounding system until you feed session logs/results back in to refine the skill, gates, and artifacts for next time.

Define gates first; they determine whether automation is safe.

For code, tests can be a deterministic gate; for product work, customer validation is the bottleneck, so you need lightweight filters (synthetic critique, mined research) before investing real customer time.

Use Git (or equivalent) to make AI iteration reversible.

Version control lets teams accept AI-driven changes confidently, compare iterations across model upgrades, and “revert forward” by adjusting the latest version rather than rolling back blindly.

Human-authored skills reduce slop and keep intent crisp.

Tyler argues first drafts of skills are often better when written by humans, because otherwise you stay on the “AI treadmill” and the skill balloons into token-heavy, overly opinionated mega-prompts.

Hooks add determinism where prompts fail.

Startup, tool, and close-session hooks can enforce rules (e.g., block dangerous shell commands or credential sharing) and trigger automatic retrospectives—things a model may forget even if instructed.

WORDS WORTH SAVING

5 quotes

We talk a lot about vibe coding, but we don't talk as much about vibe PMing.

— Tyler Folkman

For me, the difference is an agent is doing something on its own and making decisions versus you consistently prompting it.

— Tyler Folkman

This is almost always the most important part. As much as possible, being able to make this deterministic is critical.

— Tyler Folkman

I've got... I can't even tell you the number of documents shot over to me that are like 10 to 20 pages of AI, and I'm just like, "Yeah, that's not happening."

— Tyler Folkman

The customer doesn't experience your defect rate. They experience the number of defects that you push out there.

— Tyler Folkman

Loops vs prompts vs agentsHerder/terminal orchestration and multi-agent workflowFour-part loop model: fetch → work → gate → artifact (+ learning)Gates, evals, and customer validation at scaleSkills in .claude and human-authored skill designClaude hooks (startup/tool/close-session) for safety and improvementPrototype variant generation and synthetic user feedbackDocs written for AI vs docs written for humansQuality loops for engineering and defect-rate vs defect-countRise of the product builder and cross-functional upskilling

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.